Sickle Cell Disease Severity Prediction from Percoll Gradient Images using Graph Convolutional Networks
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Sickle Cell Disease Severity Prediction from
Percoll Gradient Images using Graph
Convolutional Networks
Ario Sadafi1,2 , Asya Makhro3 , Leonid Livshits3 , Nassir Navab2,4 ,
Anna Bogdanova3 , Shadi Albarqouni2,5 , and Carsten Marr1
arXiv:2109.05372v1 [eess.IV] 11 Sep 2021
1
Institute of Computational Biology, Helmholtz Zentrum München - German
Research Center for Environmental Health, Neuherberg, Germany
2
Computer Aided Medical Procedures, Technical University of Munich, Germany
3
Red Blood Cell Research Group, Institute of Veterinary Physiology, Vetsuisse
Faculty and the Zurich Center for Integrative Human Physiology, University of
Zurich, Zurich, Switzerland
4
Computer Aided Medical Procedures, Johns Hopkins University, USA
5
Helmholtz AI, Helmholtz Center Munich, Neuherberg, Germany
Abstract. Sickle cell disease (SCD) is a severe genetic hemoglobin dis-
order that results in premature destruction of red blood cells. Assessment
of the severity of the disease is a challenging task in clinical routine, since
the causes of broad variance in SCD manifestation despite the common
genetic cause remain unclear. Identification of the biomarkers that would
predict the severity grade is of importance for prognosis and assessment
of responsiveness of patients to therapy. Detection of the changes in
red blood cell (RBC) density by means of separation of Percoll density
gradient could be such marker as it allows to resolve intercellular differ-
ences and follow the most damaged dense cells prone to destruction and
vaso-occlusion. Quantification of the images obtained from the distribu-
tion of RBCs in Percoll gradient and interpretation of the obtained is
an important prerequisite for establishment of this approach. Here, we
propose a novel approach combining a graph convolutional network, a
convolutional neural network, fast Fourier transform, and recursive fea-
ture elimination to predict the severity of SCD directly from a Percoll
image. Two important but expensive laboratory blood test parameters
measurements are used for training the graph convolutional network. To
make the model independent from such tests during prediction, these two
parameters are estimated by a neural network from the Percoll image di-
rectly. On a cohort of 216 subjects, we achieve a prediction performance
that is only slightly below an approach where the groundtruth laboratory
measurements are used. Our proposed method is the first computational
approach for the difficult task of SCD severity prediction. The two-step
approach relies solely on inexpensive and simple blood analysis tools and
can have a significant impact on the patients’ survival in underdeveloped
countries where access to medical instruments and doctors is limited.
Keywords: Graph Convolutional Networks · Percoll Gradients · Sever-
ity Prediction · Sickle Cell Disease
2 A. Sadafi et al.
1 Introduction
Sickle cell disease (SCD) is a disorder caused by mutations in position 6 of β-
hemoglobin gene (hemoglobin S) with extreme variability at phenotypic level.
In some patients the disease manifestation is so mild that they remain asymp-
tomatic most of the time while others die before the age of five from several of
the severe complications associated to the SCD [11]. Individuals who have the
hemoglobin S variant are naturally protected against malaria, which has a pro-
found influence on the spread of sickle cell disease globally affecting the tropical
(African and Asian) countries the most. Many of these countries are not able to
support diagnosis and appropriate healthcare for this group of patients leading
to a drop in the life expectancy from 45-55 years in high income countries to
90% death rate before the age of 5 in low income countries [12].
Severity monitoring and prediction of the SCD is therefore an important task
along with development of new effective and inexpensive therapeutic strategies.
Changes in severity allow monitoring of the treatment efficiently and for predic-
tion and prevention of life-threatening complications in short future. To date,
there is no practical test based on red blood cells (RBCs) density separation
analysis available for prediction of the severity of the disease for a patient.
An important parameter for disease severity assessment is the percentage
of hypo- and hyperchromic cells. RBCs with a hemoglobin concentration above
410 g/l are called hyperchromic and characterized by low cellular deformability
[4] and increased probability of aggregation of the hemoglobin S which directly
associates with advanced severity and poor prognosis for the SCD patients [3,2].
In contrary, Hypochromic RBCs with low hemoglobin content are associated
with a lower probability of hemoglobin S aggregation and sickling and thus with
mild disease manifestation. Measurement of these parameters using blood smears
is laborious and time-consuming, when done manually by skilled personnel, or
rather requires expensive medical laboratory equipment, when automated.
The spleen plays an important role in clearing the blood from old, broken,
dehydrated or hyperchromic red blood cells (RBCs). A normal and functioning
spleen reduces the intravascular hemolysis of damaged cells (where cells rupture
in the blood vessels) and prevents vaso-occlusive crisis (where terminally dense
sickle cells block circulation of blood vessel leading to painful crisis) and vascular
damage in SCD patients [1]. However, fibrosis and progressive atrophy of the
spleen resulting finally in necrosis of the organ, known as autosplenectomy, which
is often observed in SCD patients with severe disease phenotype. It is a known
problem in children with SCD due to repeated splenic vaso-occlusive events in
the organ [1]. Measuring spleen size with ultrasound is a common way to evaluate
the organ’s condition in SCD.
We here propose a computational approach that circumvents expensive lab
tests and relies solely on the measurement of spleen size and a Percoll image.
Percoll images are used to assess the density of the cells and particles. After
centrifugation, several bands with different thicknesses are formed by RBCs of
similar density (see Fig. 1) holding important information about a SCD patient’s
condition. Back in 1984, Fabry et al. [5] observed a decrease in the dense fraction
SCD Severity Prediction from Percoll Gradient Images 3
of Percoll images in SCD patients suffering from painful crisis in 11 patients over
14 painful crisis image. This information can also be computationally analyzed:
Sadafi et al. [10] introduced a hybrid approach based on CNNs and features
extracted from fast Fourier transform to classify a variety of hereditary hemolytic
anemias using Percoll image data.
To predict the severity of the SCD patients, we are proposing an approach
based on graph convolutional networks (GCN) to form a population graph [8] on
our data. The similarity of the GCN edges is calculated using lab (Percentage of
hypo- and hyperchromic RBCs) and clinical data (spleen size). The spleen size
is measured using ultrasound. We propose a CNN based approach to have an
easy to access and affordable anywhere in the world way to estimate required
lab data from the Percoll image.
2 Methodology
Our proposed method, SCD-severity-GCN aims at predicting SCD severity from
cheap and easy accessible patient data and consists of the following steps: (i) The
abundance of hypo- and hyperchromic cells in the blood sample are predicted
based on a Percoll image; (ii) Relevant features are extracted from the Percoll
image using a CNN and fast Fourier transform (FFT). (iii) A similarity metric
between Percolls based on a patient’s spleen size and the predicted abundance
of hypo- and hyperchromic cells is calculated to form a population graph. Using
GCNs the SCD severity is predicted (Fig 1)
2.1 Model
Our goal is to have a model f that takes a Percoll image Pi and the spleen size
Ji of a patient sample i to return a severity grade Si :
Si = f (Pi , Ji ; θ) (1)
where θ are the model parameters that are learned by training on the dataset.
2.2 Feature extraction
For primary feature extraction the approach proposed in [10] is employed. There,
the extraction of Fourier features from the images has been demonstrated to
enhance disease classification performance on Percoll images. Accordingly, we
extract features with an AlexNet [7] architecture and combine them with features
from FFT (see Fig. 1). We obtain pretrained weights of the model fcnn−fft and
use the activations preceding the final classification layer as features for our GCN
approach.
To reduce feature dimensions, we used recursive feature elimination [6] and
a Ridge classifier as suggested by Parisot et al. [8].
xi = RFE(fcnn−fft )(Pi ) (2)
4 A. Sadafi et al.
a
Percoll image (P)
98.1%
Blood test
RFE Predict
12
SCD-severity-GCN 1.1
severity 0.9
score
Ultrasound
1 2 3 4
examination
Spleen size (J) SCD severity
b
Training Data Population graph Graph convolutional layers Severity scores
Percoll image
feature
extraction
CNN-FFT
RFE
Hb density
estimator Similarity
calculation
%Hypo Sample features
Sample similarity
%Hyper
Severity score
Spleen size 1 2 3 4
Fig. 1. Overview of the proposed SCD-severity-GCN approach. a) A Percoll image from
a conventional blood test and the spleen size obtained by ultrasound examination are
passed to our trained SCD-severity-GCN to predict a SCD severity score for the patient.
b) The SCD-severity-GCN is trained in the following way: Features from the Percoll
image are extracted with a convolutional neural network and a fast Fourier transform
(CNN-FFT [10]). Another independently trained CNN estimates the hemoglobin (Hb)
density of hypochromic (Hypo) and hyperchromic (Hyper) cells. Together with the
spleen size of the patient, a similarity measure is calculated between the nodes of a
population graph. After two layers of graph convolutions, a severity score for every
sample is predicted.
where xi is the feature vector extracted for the Percoll image Pi . Also in our
approach this step improved the convergence of the training significantly.
2.3 Graph convolution network
One of the most intuitive ways of representing populations and their similarities
is through graphs. In our approach, every Percoll image Pi is represented by
a vertex v ∈ V and the similarity between the Percoll images is modelled by
weighted edges E calculated from the expensive laboratory data (the percentages
of hypo- and hyperchromic) which are predicted and cheap clinical data (i.e.
spleen size of the patient) (see Fig 2). A population graph G = {V, E} is defined
accordingly [8].
SCD Severity Prediction from Percoll Gradient Images 5
2.4 Hemoglobin density estimation
To allow for an application of the method without expensive laboratory testing,
the percentages of hypo- and hyperchromic cells Ĥ in the blood are estimated by
a regression. A CNN fchrome is proposed for this task. The groundtruth values
H are provided for every Percoll image and are used to train the network:
N
1 X
Lchrome (γ) = (Hi − Ĥi )2 (3)
N i=1
where Ĥi = fchrome (Pi ; γ) and γ is the network parameters.
2.5 Similarity metric
Under the assumption that patients with similar features experience comparable
severity of the disease, the similarity between two samples v and w is calculated
via
E(vi , vj ) = e−(||Ĥvi −Ĥvj ||+λ[Jvi ==Jvj ]) (4)
where Ĥ is the vector of estimated percentages of hypo- and hyperchromic cells
and J is the spleen size, as above. Iverson brackets yield 1 in case of equality and
0 otherwise. Note that spleen sizes are given as discrete numbers in centimeters
(see Fig. 2), obtained in the clinic with a conventional ultrasound device. The
coefficient λ is set to weight the importance of spleen and lab measurements.
3 Experiments
3.1 Dataset
Our dataset consists of the 216 samples with Percoll images and laboratory data
(% hypo, %hyper) and clinical data (spleen size) obtained from 17 patients diag-
nosed with SCD, who participated in a clinical trial (NCT03247218) conducted
in Emek Medical Center in Afula6 . The study has been conducted in accordance
with local ethics committee guidelines and the Declaration of Helsinki. Blood
samples were acquired during pre-planned monthly visits according to the trial
protocol. For every visit the patient’s health was evaluated using blood analysis,
including RBC characteristics and measurement of hemolytic and inflammatory
markers, urine analysis and blood pressure measurements. Severity of a patient’s
condition at each measurement point was estimated using the scoring approach
proposed by Sebastiani et al. [11] with minor modifications on disease severity
score calculation. Figure 2 shows distribution of severity scores and example
samples from the dataset.
6
https://clinicaltrials.gov/ct2/show/NCT03247218
6 A. Sadafi et al.
a b Severity 1 Severity 2 Severity 3 Severity 4
Number of samples
16
37 65
98
Severity: 1 2 3 4
Spleen size (cm): 0 0 removed 3 1 removed 0 0 0 0 0 0
% Hypo: 0.6 1.7 1.7 6.3 8.2 4.5 13.4 4.2 15.7 5.3 15.3 3.3
% Hyper: 2.7 2.7 2.6 0.4 3.5 1.2 4.1 3.8 4.5 9 5.2 10.9
Fig. 2. Dataset overview. a) Pie chart shows distribution of severity values for the
patients in the dataset. b) Example images sorted according to their severity scores.
Corresponding clinical and laboratory test data for each Percoll is also demonstrated.
Spleen size of the patients with autosplenectomy and splenectomy is indicated with
removed and 0 respectively.
3.2 Implementation details
Hemoglobin density estimation: A CNN with seven convolutional layers with
ReLU activation function and max-pooling is used. After global average pooling
and two fully connected layers the output is regulated with a final ReLU. Two
dropout layers with a drop rate of 0.5 are used for regularization.
Feature extraction: The output size from CNN-FFT is 1024, which is reduced
with recursive feature elimination (RFE) [6] to 50 features. These features are
used as the final feature vector for each Percoll image.
Graph convolutional network: A population graph [8] is created based on the
defined feature vectors and similarities. We use two hidden layers in the graph
and 50 filters in the hidden layers. The dropout rate is set to 0.2. For similarity
calculation λ is set to 10.
Training: Both training procedures are carried out on a 10-fold cross valida-
tion dataset. The model fchrome estimating hemoglobin density is trained with
AMSGrad variation of Adam optimizer for 100 epochs and a learning rate of
0.0005. The graph convolutional network is trained for 300 epochs using Adam
optimizer and a learning rate of 0.01. We use the Tensorflow framework for im-
plementation and training.
Evaluation metrics: We are reporting root mean square error (RMSE) for the
regression task of hemoglobin density estimation. Accuracy, weighted F1-score
and area under ROC are reported for the severity grading as well as the area
under precision recall curve for every class. Scikit-learn [9] implementation is
used for calculation all of the metrics.
Baseline: A linear SVM [9] trained on the feature vectors is used as a baseline
for our grading approach.
SCD Severity Prediction from Percoll Gradient Images 7
3.3 Results
The dataset is divided into 10 stratified folds for patient-wise cross validation.
All of the models are independently ran on each combination of these folds.
Mean and standard deviation is reported for all of the 10 experiments.
First, the values predicted by the Hb density estimation model fchrome based
on Percoll images are compared against the actual lab tests. The root mean
square error (RMSE) of the percentage for hypochromic cells is 6.5 ± 4.0 and for
hyperchromic cells 0.90±0.12. Considering the ranges of the hypo and hyper val-
ues, which are [0.6, 37.5] and [0.2, 10.9], respectively, we consider the estimation
sufficiently good.
Next, we compare our SCD-severity-GCN approach with the following meth-
ods: (i) A linear SVM trained on the xi features vectors extracted from the
Percoll image (SVM), (ii) a linear SVM trained on xi feature vectors and the
cheap clinical ultrasound and newly proposed and time consuming groundtruth
lab information (SVM - Lab), (iii) a GCN based on randomized laboratory in-
formation (GCN - Rand), and (iv) a GCN using not the estimated, but the
actual laboratory information (GCN - Lab) as the upper limit. Table 1 shows
that our SCD-severity-GCN approach using estimated Hb densities is close to
the GCN that required hard to obtain lab data (GCN - Lab) in terms of accu-
racy, weighted F1-score and area under ROC. Since the dataset is unbalanced,
we are reporting the area under precision recall curve in Figure 3 for every class
and different approaches.
0.8 GCN-Rand
0.7 SCD-severity-GCN
0.6 GCN-Lab
0.5
AU PRC
0.4
0.3
0.2
0.1
0
1 2 3 4
Severity score
Fig. 3. Area under precision recall curve for different methods per class.
3.4 Ablation study
GCNs are sensitive to the formulation of the graph adjacency matrix based on
the pairwise similarity that is defined between the nodes. Choosing parameters
that are biologically significant and easy to obtain is crucial. To evaluate the
importance of the different clinical (spleen size) and laboratory (% of hypo- and
hyperchromic cells) information used for the formation of our GCN, we designed
8 A. Sadafi et al.
Table 1. Our proposed SCD-severity-GCN method based on GCN outperforms other
approaches. Prediction of hypo- and hyperchromic cells percentage slightly affects the
performance while providing independence from expensive cell counters.
Accuracy F1 - Score AU ROC
SVM 0.44 ± 0.07 0.28 ± 0.02 -
SVM - Lab. 0.39 ± 0.14 0.29 ± 0.08 -
GCN - Rand. 0.53 ± 0.05 0.42 ± 0.08 0.53 ± 0.20
SCD-severity-GCN 0.61 ± 0.13 0.53 ± 0.17 0.61 ± 0.25
GCN - Lab. 0.65 ± 0.15 0.59 ± 0.19 0.67 ± 0.24
a b c
Severity
1 2
3 4
Feature vectors SCD-severity-GCN GCN - Lab
Fig. 4. UMAP embedding of the (a) processed feature vectors, (b) the GCN with
predicted ones and (c) the GCN with groundtruth lab information. Clear separation
of samples with different severity scores by the proposed GCN is evident.
an ablation study and compare GCNs trained with different combinations of
these parameters. As Table 2 shows, the combination based on spleen size and
percentages of hypo- and hyperchromic RBCs yields the best result.
Table 2. Combination of different clinical and laboratory measurements results in a
slightly different shape for the GCN and thus different performance.
GCN similarity parameters Accuracy F1 - Score AU ROC
Spleen 0.45 ± 0.01 0.28 ± 0.01 0.48 ± 0.18
Spleen & Hypo 0.63 ± 0.11 0.56 ± 0.14 0.63 ± 0.23
Spleen & Hyper 0.62 ± 0.16 0.54 ± 0.21 0.60 ± 0.25
Hypo & Hyper 0.55 ± 0.08 0.46 ± 0.12 0.55 ± 0.22
Spleen & Hypo & Hyper 0.65 ± 0.15 0.59 ± 0.19 0.67 ± 0.24
3.5 Discussion
Severity prediction of SCD is a challenging task normally preformed with sev-
eral clinical and laboratory tests. Here we propose a novel potential severity
prediction approach based on RBC density separation (as provided by Percoll
gradients) analysis that may amend the currently existing ones. Information
SCD Severity Prediction from Percoll Gradient Images 9
obtained solely from Percoll images is not be sufficient for an acceptable classifi-
cation (see Table 1), even though those features sufficed for successful diagnosis
of different anemias [10]. By combining Percoll derived features with comple-
mentary clinical and laboratory data and training a GCN with this information,
we can achieve an accuracy that is surprisingly high for this challenging clinical
task. This is illustrated by the UMAP embedding of feature vectors (Fig 4 a),
and GCN outputs with estimated (Fig 4 b) and groundtruth lab information
(Fig 4 c). Samples from different severity classes are nicely disentangled in the
UMAP thanks to the GCN approach we utilized. Although clustering using the
groundtruth lab information (GCN - Lab) is a lot better, a smooth transition
from low to high severity is already evident in the approach that uses estimated
Hb density only (SCD-severity-GCN).
4 Conclusion
Sickle cell disease severity prediction is an important task that allows to prevent
life-threatening complications, reduce morbidity and mortality and refine the
choice of optimal therapeutic strategies [11]. Offering affordable and versatile
solutions for improving life quality of the SCD patients is a necessity specially
in low resource areas of the planet. Here, we proposed the first computational
method requiring only the Percoll gradient image and spleen size obtained from
a conventional ultrasound. Analysis of Percoll gradient images with CNNs nicely
predicted percentages of hypo- and hyperchromic cells and the proposed GCN
predicted SCD severity score with a surprisingly high accuracy. Our approach
uses a unique combination of methods, with a GCN at its heart.
Results look very promising and provide a solid ground for future work. Next
we will analyze more patients, especially more severe ones as well as pediatric
datasets. Our SCD-severity-GCN based on Percoll images requires much smaller
volumes of blood compared to common hematological tests (1 ml or less instead
of 7-10 ml), which is particularly relevant for kids and patients suffering from
severe anemia.
Acknowledgments
Special thanks to Prof. Ariel Koren and Dr. Carina Levin from the Emek Medical
Center in Afula who made this work possible. This project has received funding
from the European Union’s Horizon 2020 research and innovation programme
under grant agreement No 675115 — RELEVANCE — H2020-MSCA-ITN-2015/
H2020-MSCA-ITN-2015. The work of L.L. was funded by UZH Foundation. C.M.
and A.S. have received funding from the European Research Council (ERC)
under the European Union’s Horizon 2020 research and innovation programme
(Grant agreement No. 866411).
10 A. Sadafi et al.
References
1. Brousse, V., Buffet, P., Rees, D.: The spleen and sickle cell disease: the sick (led)
spleen. British journal of haematology 166(2), 165–176 (2014)
2. Brugnara, C., Mohandas, N.: Red cell indices in classification and treatment of
anemias: from mm wintrobes’s original 1934 classification to the third millennium.
Current opinion in hematology 20(3), 222–230 (2013)
3. De Franceschi, L.: Pathophisiology of sickle cell disease and new drugs for the treat-
ment. Mediterranean Journal of Hematology and Infectious Diseases 1(1) (2009)
4. Deuel, J.W., Lutz, H.U., Misselwitz, B., Goede, J.S.: Asymptomatic elevation of
the hyperchromic red blood cell subpopulation is associated with decreased red
cell deformability. Annals of hematology 91(9), 1427–1434 (2012)
5. Fabry, M.E., Benjamin, L., Lawrence, C., Nagel, R.L.: An objective sign in painful
crisis in sickle cell anemia: the concomitant reduction of high density red cells.
Blood 64(2), 559–563 (1984)
6. Guyon, I., Weston, J., Barnhill, S., Vapnik, V.: Gene selection for cancer classifi-
cation using support vector machines. Machine learning 46(1), 389–422 (2002)
7. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep con-
volutional neural networks. Advances in neural information processing systems 25,
1097–1105 (2012)
8. Parisot, S., Ktena, S.I., Ferrante, E., Lee, M., Moreno, R.G., Glocker, B., Rueck-
ert, D.: Spectral graph convolutions for population-based disease prediction. In:
International conference on medical image computing and computer-assisted in-
tervention. pp. 177–185. Springer (2017)
9. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O.,
Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine
learning in Python. Journal of Machine Learning Research 12, 2825–2830 (2011)
10. Sadafi, A., Moya Sans, L.M., Makhro, A., Livshits, L., Navab, N., Bogdanova, A.,
Albarqouni, S., Marr, C.: Fourier transform of percoll gradients boosts cnn classi-
fication of hereditary hemolytic anemias. In: 2021 IEEE International Symposium
on Biomedical Imaging (ISBI). IEEE (2021)
11. Sebastiani, P., Nolan, V.G., Baldwin, C.T., Abad-Grau, M.M., Wang, L., Adewoye,
A.H., McMahon, L.C., Farrer, L.A., Taylor, J.G., Kato, G.J., et al.: A network
model to predict the risk of death in sickle cell disease. Blood 110(7), 2727–2735
(2007)
12. Wastnedge, E., Waters, D., Patel, S., Morrison, K., Goh, M.Y., Adeloye, D., Rudan,
I.: The global burden of sickle cell disease in children under five years of age: a
systematic review and meta-analysis. Journal of global health 8(2) (2018)You can also read
